search

nicholasjclark / mvgam

{mvgam} R 📦 to fit Dynamic Bayesian Generalized Additive Models for time series analysis and forecasting
https://nicholasjclark.github.io/mvgam/
Other
91 stars 11 forks source link

Add pathfinder support #37

Closed nicholasjclark closed 5 months ago

nicholasjclark commented 7 months ago

New versions of Cmdstan and cmdstanr allow support for the pathfinder algorithm, which employs Quasi-Newton Variational Inference (https://www.youtube.com/watch?v=TPptuDp-w2E). Requires Cmdstan version 2.33 or higher, as well as cmdstanr version 0.6.1.9000 or higher (this currently only applies to the GitHub version). Will need to check that the resulting object of class CmdStanPathfinder can still be used in the same way as the variational object (class CmdStanVB). Use ??cmdstanr::model to find details

nicholasjclark commented 5 months ago

Done, though this doesn't seem to work too well with all of mvgam's latent parameters